sensitivity-analysis
Purpose
Quantify which parameters, assumptions, interactions, or uncertainties dominate a conclusion.
Input contract
mode_contracts:
Morris: &sensitivity_input
required: [model_or_reasoning_chain, uncertain_parameters, conclusion_metric]
optional: [input_distributions, perturbation_bounds, decision_options]
constraints: [ranges_and_dependencies_must_be_declared]
Sobol: *sensitivity_input
perturbation: *sensitivity_input
Monte-Carlo: *sensitivity_input
Execution protocol
Do not perform called SOP operations inline; each loaded SOP owns its contract and thresholds.
- You MUST load skill
define-analysis-dimensions to define the analysis dimensions.
- Apply the selected mode's perturbation, decomposition, or propagation operation.
- You MUST load skill
assess-sensitivity to estimate the declared sensitivity effects.
- You MUST load skill
identify-load-bearing-factors to identify load-bearing factors.
- You MUST load skill
quantify-information-value to quantify information value for unresolved drivers.
If the dominant sensitivity is caused by an unstable problem frame rather than an input value, consider problem-reframing as the next tactic.
Deviation: local-perturbation, Morris, Sobol, Monte-Carlo, and decision-value modes select subsets, but mode and omitted analyses must be recorded.
Mode branches
Morris: screen many uncertain inputs with elementary effects to identify influential factors and interactions before expensive global analysis. You MUST load skill apply-perturbation to generate elementary-effect trajectories. You MUST load skill decompose-global-sensitivity to separate influential factors and interactions.
Sobol: decompose output variance into first-order and total-order contributions when the model can support global sampling. You MUST load skill decompose-global-sensitivity to compute the variance decomposition. You MUST load skill propagate-uncertainty to propagate the declared input distributions.
perturbation: vary declared inputs around a baseline to expose local directional sensitivity and threshold crossings. You MUST load skill apply-perturbation to execute the bounded variations.
Monte-Carlo: propagate input distributions through repeated draws to quantify outcome uncertainty rather than relying on a point estimate. You MUST load skill propagate-uncertainty to run and summarize the repeated draws.
Output contract
mode_contracts:
Morris: &screening_sensitivity_output
produces: [sensitivity_profile, interaction_effects]
delta_fields: [findings, evidence_updates, uncertainties, decisions, recommended_jumps]
Sobol:
produces: [sensitivity_profile, interaction_effects, uncertainty_contributions]
delta_fields: [findings, evidence_updates, uncertainties, decisions, recommended_jumps]
perturbation: *screening_sensitivity_output
Monte-Carlo:
produces: [sensitivity_profile, uncertainty_contributions, information_value_ranking]
delta_fields: [findings, evidence_updates, uncertainties, decisions, recommended_jumps]
Thresholds and quality gates
- B: dominant drivers, interactions, ranges, and uncertainty contributors are separated; local effects are not presented as global effects.
Failure and counterexamples
Reject rankings based on arbitrary ranges, invalid model runs, or point estimates that hide distributional uncertainty.
Provenance map
sensitivity-analysis: intermediate (exact pool names are package-prefixed variants).
parameter-screening, variance-decomposition, assumption-criticality, uncertainty-propagation, decision-sensitivity, screening-then-decomposition, uncertainty-cascade: resolved.
Preserved source criteria ledger
- Preserve perturbation, Morris/Sobol, uncertainty cascade, assumption criticality, and information-value modes.
Context checkpoint / Delta notes
Append dimensions, perturbation results, driver rankings, uncertainty updates, and next evidence decisions.
1---2name: sensitivity-analysis3description: "Quantify which parameters, assumptions, interactions, or uncertainties dominate the conclusion; select Morris/Sobol/perturbation/Monte-Carlo as modes."4metadata:5 source: https://skillmd.com/skills/yogsoth-ai/sensitivity-analysis6---78# sensitivity-analysis9## Purpose10Quantify which parameters, assumptions, interactions, or uncertainties dominate a conclusion.11## Input contract12```yaml13mode_contracts:14 Morris: &sensitivity_input15 required: [model_or_reasoning_chain, uncertain_parameters, conclusion_metric]16 optional: [input_distributions, perturbation_bounds, decision_options]17 constraints: [ranges_and_dependencies_must_be_declared]18 Sobol: *sensitivity_input19 perturbation: *sensitivity_input20 Monte-Carlo: *sensitivity_input21```22## Execution protocol23Do not perform called SOP operations inline; each loaded SOP owns its contract and thresholds.24251. You MUST load skill `define-analysis-dimensions` to define the analysis dimensions.262. Apply the selected mode's perturbation, decomposition, or propagation operation.273. You MUST load skill `assess-sensitivity` to estimate the declared sensitivity effects.284. You MUST load skill `identify-load-bearing-factors` to identify load-bearing factors.295. You MUST load skill `quantify-information-value` to quantify information value for unresolved drivers.30 If the dominant sensitivity is caused by an unstable problem frame rather than an input value, consider `problem-reframing` as the next tactic.31Deviation: `local-perturbation`, `Morris`, `Sobol`, `Monte-Carlo`, and decision-value modes select subsets, but mode and omitted analyses must be recorded.3233## Mode branches3435- `Morris`: screen many uncertain inputs with elementary effects to identify influential factors and interactions before expensive global analysis. You MUST load skill `apply-perturbation` to generate elementary-effect trajectories. You MUST load skill `decompose-global-sensitivity` to separate influential factors and interactions.36- `Sobol`: decompose output variance into first-order and total-order contributions when the model can support global sampling. You MUST load skill `decompose-global-sensitivity` to compute the variance decomposition. You MUST load skill `propagate-uncertainty` to propagate the declared input distributions.37- `perturbation`: vary declared inputs around a baseline to expose local directional sensitivity and threshold crossings. You MUST load skill `apply-perturbation` to execute the bounded variations.38- `Monte-Carlo`: propagate input distributions through repeated draws to quantify outcome uncertainty rather than relying on a point estimate. You MUST load skill `propagate-uncertainty` to run and summarize the repeated draws.3940## Output contract41```yaml42mode_contracts:43 Morris: &screening_sensitivity_output44 produces: [sensitivity_profile, interaction_effects]45 delta_fields: [findings, evidence_updates, uncertainties, decisions, recommended_jumps]46 Sobol:47 produces: [sensitivity_profile, interaction_effects, uncertainty_contributions]48 delta_fields: [findings, evidence_updates, uncertainties, decisions, recommended_jumps]49 perturbation: *screening_sensitivity_output50 Monte-Carlo:51 produces: [sensitivity_profile, uncertainty_contributions, information_value_ranking]52 delta_fields: [findings, evidence_updates, uncertainties, decisions, recommended_jumps]53```54## Thresholds and quality gates55- B: dominant drivers, interactions, ranges, and uncertainty contributors are separated; local effects are not presented as global effects.56## Failure and counterexamples57Reject rankings based on arbitrary ranges, invalid model runs, or point estimates that hide distributional uncertainty.58## Provenance map59- `sensitivity-analysis`: intermediate (exact pool names are package-prefixed variants).60- `parameter-screening`, `variance-decomposition`, `assumption-criticality`, `uncertainty-propagation`, `decision-sensitivity`, `screening-then-decomposition`, `uncertainty-cascade`: resolved.61## Preserved source criteria ledger62- Preserve perturbation, Morris/Sobol, uncertainty cascade, assumption criticality, and information-value modes.63## Context checkpoint / Delta notes64Append dimensions, perturbation results, driver rankings, uncertainty updates, and next evidence decisions.